In [1]:
# Install a pip package in the current Jupyter kernel
import sys
!{sys.executable} -m pip install numpy==1.16.4
In [2]:
!conda list
In [2]:
import os
import numpy as np
import cv2
from PIL import Image
from keras_unet.utils import plot_imgs, get_augmented
from sklearn.model_selection import train_test_split
from keras_unet.models import custom_unet
from keras.callbacks import ModelCheckpoint, EarlyStopping
from keras.optimizers import Adam, SGD
from keras_unet.metrics import iou, iou_thresholded
from keras.callbacks import ModelCheckpoint
In [3]:
IMG_TRAIN_PATH = 'data/train_images/'
MSK_TRAIN_PATH = 'data/train_masks/'
IMG_TEST_PATH = 'data/test_images/'
MSK_TEST_PATH = 'data/test_masks/'
IMG_VAL_PATH = 'data/val_images/'
MSK_VAL_PATH = 'data/val_masks/'
In [4]:
ALL_PATHS = [IMG_TRAIN_PATH, MSK_TRAIN_PATH, IMG_TEST_PATH,
MSK_TEST_PATH, IMG_VAL_PATH, MSK_VAL_PATH]
In [5]:
all_training_images = [f for f in os.listdir(IMG_TRAIN_PATH) if f.endswith('.png')]
all_training_masks = [f for f in os.listdir(MSK_TRAIN_PATH) if f.endswith('.png')]
In [12]:
imgs_list = []
masks_list = []
for image, mask in zip(all_training_images, all_training_masks):
imgs_list.append(np.array(Image.open(IMG_TRAIN_PATH + image).resize((256,256))))
msk = cv2.imread(MSK_TRAIN_PATH + image, cv2.IMREAD_GRAYSCALE)
msk = cv2.resize(msk, (256, 256), interpolation=cv2.INTER_NEAREST)
masks_list.append(msk)
imgs_np = np.asarray(imgs_list)
masks_np = np.asarray(masks_list)
In [13]:
X = np.asarray(imgs_np, dtype=np.float32)/255
y = np.asarray(masks_np, dtype=np.float32)
In [14]:
plot_imgs(org_imgs=X, mask_imgs=y, nm_img_to_plot=10, figsize=6)
In [15]:
y = np.where(y == 0, 0, 1)
In [16]:
plot_imgs(org_imgs=X, mask_imgs=y, nm_img_to_plot=10, figsize=6)
In [17]:
y = y.reshape(y.shape[0], y.shape[1], y.shape[2], 1)
In [18]:
X_train, X_val, y_train, y_val = train_test_split(X, y, test_size=0.2, random_state=0)
print("x_train: ", X_train.shape)
print("y_train: ", y_train.shape)
print("x_val: ", X_val.shape)
print("y_val: ", y_val.shape)
In [19]:
train_gen = get_augmented(
X_train, y_train, batch_size=32,
data_gen_args = dict(
rotation_range=10.,
#width_shift_range=0.02,
height_shift_range=0.02,
#shear_range=5.,
zoom_range=0.2,
horizontal_flip=True,
vertical_flip=False,
fill_mode='constant'
))
In [15]:
unique, counts = np.unique(y, return_counts=True)
In [16]:
for i in range(0,len(counts)):
a = counts[i]/counts.sum()
print(str(a))
In [20]:
INPUT_SHAPE = X_train[0].shape
BATCH_SIZE = 32
In [21]:
model = custom_unet(
input_shape = INPUT_SHAPE,
filters = 32,
use_batch_norm = True,
dropout = 0.3,
dropout_change_per_layer=0.0,
num_layers=4)
#output_activation = 'softmax')
In [27]:
model.summary()
In [50]:
model_filename = 'segm_model_v1.h5'
early_stopper = EarlyStopping(patience = 5, verbose = 1)
callback_checkpoint = ModelCheckpoint(
model_filename,
verbose=1,
monitor='val_loss',
save_best_only=True,
)
In [53]:
def focal_loss(target, output, gamma=2):
output /= K.sum(output, axis=-1, keepdims=True)
eps = K.epsilon()
output = K.clip(output, eps, 1. - eps)
return -K.sum(K.pow(1. - output, gamma) * target * K.log(output),
axis=-1)
In [24]:
import keras.backend as K
import tensorflow as tf
from keras_unet.losses import jaccard_distance
In [26]:
model.compile(
optimizer='adam',
loss=jaccard_distance,
#loss = 'binary_crossentropy',
metrics=['accuracy',iou,iou_thresholded])
In [28]:
steps_per_epoch = len(X_train) // 32
In [30]:
history = model.fit_generator(
train_gen,
steps_per_epoch=steps_per_epoch,
epochs=10,
validation_data=(X_val, y_val),
callbacks=[early_stopper, callback_checkpoint])
In [60]:
model_filename = 'segm_model_v1_c2.h5'
model.load_weights(model_filename)
# y_pred = model.predict(x_val)
In [36]:
all_testing_images = [f for f in os.listdir(IMG_TEST_PATH) if f.endswith('.png')]
all_testing_masks = [f for f in os.listdir(MSK_TEST_PATH) if f.endswith('.png')]
In [37]:
imgs_list = []
masks_list = []
for image, mask in zip(all_testing_images, all_testing_masks):
imgs_list.append(np.array(Image.open(IMG_TEST_PATH + image).resize((256,256))))
msk = cv2.imread(MSK_TEST_PATH + image, cv2.IMREAD_GRAYSCALE)
msk = cv2.resize(msk, (256, 256), interpolation=cv2.INTER_NEAREST)
masks_list.append(msk)
imgs_np = np.asarray(imgs_list)
masks_np = np.asarray(masks_list)
In [38]:
masks_np = np.where(masks_np > 0, 1, 0)
In [39]:
X_test = np.asarray(imgs_np, dtype=np.float32)/255
y_test = np.asarray(masks_np, dtype=np.float32)
In [61]:
y_pred = model.predict(X_val, verbose=1)
In [62]:
from keras_unet.utils import plot_imgs
plot_imgs(org_imgs=X_val, mask_imgs=y_val, pred_imgs=y_pred, nm_img_to_plot=10)
